Prerequisites

• Node.js 18+ (LTS recommended)
• npm 9+
• Rust stable toolchain
• Platform requirements for Tauri 2 (macOS/Linux/Windows build dependencies)

Build & Verification

npm run build
cd src-tauri && cargo test

Additional project checks:

npm run check:translation-linking
npm run check:prod-no-testids

Trust, Data & Installation

• Lumenfolio is local-first. Indexes, notes, chat history and translation metadata are stored locally.
• Notes are additionally written as plain .md files in the folder you choose, and the database can be snapshotted to a folder you choose.
• API keys are currently stored locally; migration to the system keychain is planned.
• If a cloud chat or translation provider is configured, selected text, questions, page context or translation content may be sent to that provider.
• If you choose a local Codex / Claude Code provider, questions, conversation memory and retrieved evidence are passed to that local CLI; the model request is handled by the CLI and its signed-in account.
• macOS builds are currently ad-hoc signed. Developer ID signing and notarization are planned.
• Release assets include SHA-256 checksums plus license, notice, AGPL sidecar license, and PDFMathTranslate source archive.

Quick Start

npm install
npm run tauri:dev

For browser-only UI iteration:

npm run dev

Trending Papers (optional)

An optional, local-first discovery feed of trending papers from Hugging Face, with Daily/Weekly/Monthly scopes and one-click add into your library. Nothing is fetched until you open it, and a PDF is downloaded only on an explicit add. It is a side utility, not the centre of the app — turn it off in Settings and the entry disappears.

Knowledge API for External Tools

Your library does not have to be reachable only from inside this app. Turn on the Knowledge API in Settings and Lumenfolio runs a local, read-only MCP endpoint that any MCP client — Claude Code, Codex, your own harness — can query.

claude mcp add --transport http lumenfolio http://127.0.0.1:37650/mcp \
  --header "Authorization: Bearer <token>"

Settings has one-click commands for Claude Code and Codex with the token filled in, so connecting is a copy and a paste.

The external client gets the same retrieval tools the in-app agent uses: list_sources to see what the library holds, search_library_knowledge to find sources by topic, then search_chunks, open_pages, read_sheet, inspect_tables and the rest to read one — all returning quotes with their source, page and bounding box.

Two differences from the in-app path, both deliberate:

- An unknown source id is an error, not a substitution. In-app, a model that invents an id gets the open document, which is the right degradation mid-turn. A program that names a source must be told it does not exist, rather than handed a different one's text and no way to notice. - Web tools are off. An external harness has its own network access; proxying its requests would lend it this app's proxy and API keys as an egress path.

The boundary this crosses. It is off by default. While on, it binds loopback only and never listens on the network — but any process on your machine holding the token can read your entire knowledge base without going through the app. The token can be rotated at any time, which restarts the endpoint and cuts off anything still using the old one. It is read-only: nothing reachable through it can create, edit or delete.